RETAIL FORECASTING: RESEARCH AND PRACTICE - MUNICH PERSONAL REPEC ARCHIVE

Page created by Lee Leon
 
CONTINUE READING
RETAIL FORECASTING: RESEARCH AND PRACTICE - MUNICH PERSONAL REPEC ARCHIVE
Munich Personal RePEc Archive

Retail forecasting: research and practice

Fildes, Robert and Ma, Shaohui and Kolassa, Stephan

Lancaster University Management School, UK, School of Business,
Nanjing Audit University, China, SAP, Switzerland

October 2019

Online at https://mpra.ub.uni-muenchen.de/89356/
MPRA Paper No. 89356, posted 19 Nov 2018 16:06 UTC
RETAIL FORECASTING: RESEARCH AND PRACTICE - MUNICH PERSONAL REPEC ARCHIVE
Management Science
                            Working Paper 2018:04

       Retail forecasting: research and practice

   Robert Fildes, Lancaster Centre for Marketing Analytics and
    Forecasting, Lancaster University Management School, UK
Shaohui Ma, School of Business, Nanjing Audit University, China
                      Stephan Kolassa, SAP Switzerland

                    The Department of Management Science
                    Lancaster University Management School
                              Lancaster LA1 4YX
                                      UK

                       © Robert Fildes, Shaohui Ma, Stephan Kolassa
                   All rights reserved. Short sections of text, not to exceed
                 two paragraphs, may be quoted without explicit permission,
                         provided that full acknowledgment is given.

                      LUMS home page: http://www.lums.lancs.ac.uk.
Centre home page: http://www.lancaster.ac.uk/lums/research/research-centres--areas/centre-for-
                           marketing-analytics-and-forecasting/

1. R.Fildes@lancaster.ac.uk 2. shaohui.ma@hotmail.com 3. stephan.kolassa@sap.com
Retail forecasting: research and practice
              Robert Fildes1, Lancaster Centre for Marketing Analytics and Forecasting
                  Department of Management Science, Lancaster University, LA1 1
            Shaohui Ma2, School of Business, Nanjing Audit University, Nanjing, 211815, China
                          Stephan Kolassa3, SAP Switzerland, SAP Switzerland
                                      8274 Tägerwilen, Switzerland

Abstract
This paper first introduces the forecasting problems faced by large retailers, from the strategic to the
operational, from the store to the competing channels of distribution as sales are aggregated over
products to brands to categories and to the company overall. Aggregated forecasting that supports
strategic decisions is discussed on three levels: the aggregate retail sales in a market, in a chain, and in
a store. Product level forecasts usually relate to operational decisions where the hierarchy of sales data
across time, product and the supply chain is examined. Various characteristics and the influential
factors which affect product level retail sales are discussed. The data rich environment at lower
product hierarchies makes data pooling an often appropriate strategy to improve forecasts, but success
depends on the data characteristics and common factors influencing sales and potential demand.
Marketing mix and promotions pose an important challenge, both to the researcher and the practicing
forecaster. Online review information too adds further complexity so that forecasters potentially face a
dimensionality problem of too many variables and too little data. The paper goes on to examine
evidence on the alternative methods used to forecast product sales and their comparative forecasting
accuracy. Many of the complex methods proposed have provided very little evidence to convince as to
their value, which poses further research questions. In contrast, some ambitious econometric methods
have been shown to outperform all the simpler alternatives including those used in practice. New
product forecasting methods are examined separately where limited evidence is available as to how
effective the various approaches are. The paper concludes with some evidence describing company
forecasting practice, offering conclusions as to the research gaps but also the barriers to improved
practice.

Keywords; retail forecasting; product hierarchies; big data; marketing analytics; user-generated web
content; new products; comparative accuracy; forecasting practice.

    1. R.Fildes@lancaster.ac.uk 2. shaohui.ma@hotmail.com 3. stephan.kolassa@sap.com
1. Introduction

     The retail industry is experiencing rapid developments both in structure, with the growth

in on-line business, and in the competitive environment which companies are facing. There is

no simple story that transcends national boundaries, with different national consumers behaving

in very different ways. For example, in 2017 on-line retailing accounted for 14.8 % of retail

sales in the US, 17.6% in the UK but only 3.4% in Italy contrasting with Germany showing a

3.5% increase to 15.1% since 2015 (www.retailresearch.org/onlineretailing.php). But whatever

the retailer’s problem, its solution will depend in part on demand forecasts, delivered through

methods and processes embedded in a forecasting support system (FSS). High accuracy

demand forecasting has an impact on organizational performance because it improves many

features of the retail supply chain. At the organizational level, sales forecasts are essential inputs

to many decision activities in functional areas such as marketing, sales, and

production/purchasing, as well as finance and accounting. Sales forecasts also provide the basis

for national, regional and local distribution and replenishment plans.

     Much effort has been devoted over the past several decades to the development and

improvement of forecasting models. In this paper we review the research as it applies to retail

forecasting, drawing boundaries around the field to focus on food, non-food including electrical

goods (but excluding for example, cars, petrol or telephony), and non-store sales (catalog and

now internet). This broadly matches the definitions and categories adopted, for example, in the

UK and US government retail statistics. Our objective is to draw together and critically evaluate

a diverse research literature in the context of the practical decisions that retailers must make

that depend on quantitative forecasts. In this examination we look at the variety of demand

patterns in the different marketing contexts and levels of aggregation where forecasts must be

made to support decisions, from the strategic to the operational. Perhaps surprisingly, given the

importance of retail forecasting, we find the research literature is both limited and often fails to

address the retailer’s decision context.

     In the next section we consider the decisions retailers make, from the strategic to the

operational, and the different levels of aggregation from the store up to the retail chain. Section

                                                  1
three considers aggregate forecasting from the market as a whole where, as we have noted,

rapid changes are taking place, down to the individual store where again the question of where

stores should be located has risen to prominence with the changes seen in shopping behavior.

We next turn to more detailed Stock Keeping Unit (SKU) forecasting, and the hierarchies these

SKUs naturally fall into. The data issues faced when forecasting include stock-outs, seasonality

and calendar events while key demand drivers are the marketing mix and promotions. On-line

product reviews and social media are new information sources that requires considerable care

if they are to prove valuable in forecasting. Section 5 provides an evaluation of the different

models used in product level demand forecasting in an attempt to provide definitive evidence

as to the circumstances where more complex methods add value. New product forecasting

requires different approaches and these are considered in Section 6. Practice varies dramatically

across the retail sector, in part because of its diversity, and in Section 7 we provide various

vignettes based on case observation which capture some of the issues retailers face and how

they provide operational solutions. Finally, Section 8 contains our conclusions as to those areas

where evidence is strong as to best practice and where research is most needed.

2. Retailers’ forecasting needs

Strategic level

    Retailers like all commercial organizations must make decisions as to their strategic

development within a changing competitive and technological environment. The standard

elements defining a retail strategy embracing market and competitive factors within the

developing technological and regulatory environment (see, for example, Levy, Weitz, and

Grewal, 2012) are typically dependent on forecasts. Fig.1 illustrates these issues showing the

recent growth of on-line purchases in the US, UK and Europe, with some suggestion that those

countries with the highest penetration levels are seeing a slowing of growth (but with clear

differences between countries and cultures). Also shown is a naïve extrapolation for 2020 using

the average growth rate from 2014 to 2016. Fig.2 shows the changing share of low-price

retailers in the UK and the US from 1994 to 2017 with forecasts to 2020 compared to the

established leaders (produced via ETS). These simple extrapolative forecasts highlight the

                                               2
strategic threat on-line and low price retailers pose, exacerbated by a dominant player in

Amazon.

                                        Fig. 1 Online shares of Retail Trade

                 (Source: Center for Retail Research: www.retailresearch.org/onlineretailing.php)

Fig. 2 Share of grocery retailers compared to the low price retailers (Aldi and Lidl) in the UK, 1994 to 2017 with ETS

                  forecasts to 2020. Source: http://www.fooddeserts.org/images/supshare.htm.

                                                          3
These figures and the extrapolative forecasts show the rapid changes in the retail

environment which require companies to respond. For example, a channel decision to develop

an on-line presence will depend on a forecast time horizon looking decades ahead but with

some quantitative precision required over shorter horizons, perhaps as soon as its possible

implementation a year or more ahead. The retailer chain’s chosen strategy will require decisions

that respond to the above changes: on location including channels, price/quality position and

target market segment(s), store type (in town vs megastores) and distribution network. A key

point is that such decisions will all typically have long-term consequences with high costs

incurred if subsequent changes are needed, flexibility being low (e.g. site location and the move

to more frequent local shopping in the UK, away from the large out-of-town stores, leading

Tesco in 2015 to sell 14 of its earmarked sites in the UK and close down others and, in 2018,

M&S proposing to close down more than 10% of its stores). Strategic forecasts are therefore

required at both at a highly aggregate level and also a geographic specific level over a long

forecast horizon.

     The small local retailer faces just as volatile an environment, with uncertainty as to the

location and target market (and product mix). Some compete directly with national chains

where the issue is what market share can be captured and sustained. But while many of the

questions faced by the national retailers remain relevant (e.g. on-line offering) there is little in

the research literature that is even descriptive of the results of the many small shop location

decisions. Exceptions include charity shops (Alexander, Cryer, and Wood, 2008) and

convenience stores (Wood and Browne, 2007) while a number of studies examine restaurants

which are outside our scope. But in this article, we focus on larger retailers carrying a wide

range of products.

Tactical level

     Tactical decisions necessarily fit within the strategic framework developed above. But

these strategic decisions do not determine the communications and advertising plan for the

chain, the categories of products to be offered, nor the variety (range) of products within each

category. At the chain level, the aim is to maximize overall profitability using both advertising

(at chain and store level) and promotional tools to achieve success.

                                                 4
At the category level the objective again is to maximize category (rather than brand) profits

which will require a pricing/ promotional plan that determines such aspects as the number and

depth of promotions over the planning horizon (of perhaps a year), their frequency, and whether

there are associated display and feature advertising campaigns. These plans are in principle

linked to operational promotional pricing decisions discussed below. The on-shelf availability

of products is also a key metric of retail service, and this depends crucially on establishing a

relationship between the product demand forecasts, inventory investment and the distribution

system. The range of products listed raises the question of new product introduction into a

category, the expected sales and its effect on sales overall (particularly within category).

     Demands placed on the warehouse and distribution system by store × product demand also

need forecasting. This is needed to plan the workforce where the number and ‘size’ of products

determines the pick rate which in turn determines the workforce and its schedule. The

constitution of the delivery fleet and planned routes similarly depend on store demand forecasts

(somewhat disaggregated) since seasonal patterns of purchasing vary by region. This is true

whether the retailer runs its own distribution network or has it outsourced to a service provider

– or, what is most common, uses a mixture, with many products supplied from the retailer’s

own distribution centers, but others supplied directly by manufacturers to stores (Direct Store

Delivery).

Operational level

     To be successful in strategic and tactical decisions, the retail company needs to design its

demand and supply planning processes to avoid customer service issues along with

unnecessarily high inventory and substantial write off costs due to obsolete products. These are

sensitive issues in retail companies because of the complexity in the demand data with

considerable fluctuations, the presence of many intermediaries in the process, diversity of

products and the service quality required by the consumer. In a general way, accurate demand

forecasting is crucial in organizing and planning purchasing, distribution, and the labor force,

as well as after-sales services. Therefore, the ability of retail managers to estimate the probable

sales quantity at the SKU × store level over the short-term leads to improved customer

satisfaction, reduced waste, increased sales revenue and more effective and efficient

                                                5
distribution.

     As a result of these various operational decisions with their financial consequences, the

cash retailers generate (since suppliers are usually paid in arrears) leads to a cash management

investment problem. Thus the cash available for investment, itself dependent on the customer

payment arrangements, needs to be forecast.

     Day-to-day store operations are also forecast dependent. In particular, staffing schedules

depend on anticipated customer activity and product intake.

3. Aggregate retail sales forecasting

     All forecasting in retail depends on a degree of aggregation. The aggregations could be on

product units, location or time buckets or promotion according to the objective of the

forecasting activity.

                        Fig. 3 Hierarchy of aggregate retail sales forecasting

     In this section, the aggregate retail sales forecasting refers to the total retail sales in a

market, a chain, or a store, as opposed to product (SKU/brand/category) specific forecasts,

i.e., we implicitly aggregate across products and promotions and up to a specific granularity

(e.g., weekly or monthly) in the time dimension, see Fig.3. Aggregate retail sales are usually

measured as a dollar amount instead of units of the products. We below review the existing

researches on three levels separately: the aggregate retail sales in a market, in a chain, and in a

                                                  6
store. Though forecasting aggregate sales at these three levels share many common issues,

e.g., seasonality and trend, they raise different forecasting questions; have different

objectives, data characteristics, and solutions.

3.1 Market level aggregate sales forecasting

            Market level aggregate sales forecasting concerns the forecasts of total sales of a retail

format, section, or the whole industry in a country or region. The time bucket for the market

level forecasts may be monthly, quarterly or yearly. The forecasts of market level retail sales

are necessary for (large) retailers both to understand changing market conditions and how these

affect their own total sales (Alon, Qi, and Sadowski, 2001). They are also central to the planning

and operation of a retail business at the strategic chain level in that they help identify the growth

potential of different business modes and stimulate the development of new strategies to

maintain market position.

        Market level aggregate retail sales data often exhibit strong trend, seasonal variations,

serial correlation and regime shifts because any long span in the data may include both

economic growth, inflation and unexpected events (Fig. 4). Time series models have provided

a solution to capturing these stylized characteristics. Thus, time series models have long been

applied for market level aggregate retail sales forecasting (e.g., Alon et al., 2001; Bechter and

Rutner, 1978; Schmidt, 1979; Zhang and Qi, 2005). Simple exponential smoothing and its

extensions to include trend and seasonal (Holt-Winters), and ARIMA models have been the

most frequent time series models employed for market level sales forecasting. Even in the

earliest references, reflecting controversies in the macroeconomic literature, the researchers

raised the question of which of various time series models performed best and how they

compared with simple econometric models 1 . The early studies suffered from a common

weakness – a failure to compare models convincingly.

1   Typically, macro econometric models do not include retail sales as an endogenous variable but rather use a
variable such as consumption.

                                                           7
300000
                             250000
         US monthly retail

                             200000
                             150000

                                           1995            2000           2005           2010     2015

                                                                        Time

                                      Fig. 4 US retail sales monthly series in million dollars.

                                                   (Source: U.S. Census Bureau)

    Some researchers found that standard time series models were sometimes inadequate to

approximate aggregate retail sales, identifying evidence of nonlinearity and volatility in the

market level retail sales time series. Thus, researchers have resorted to nonlinear models,

especially artificial neural networks (Alon, et al., 2001; Chu and Zhang, 2003; Zhang and Qi,

2005). Results have indicated that traditional time series models with stochastic trend, such as

Winters exponential smoothing and ARIMA, performed well when macroeconomic conditions

were relatively stable. When economic conditions were volatile (with rapid changes in

economic conditions) ANNs was claimed to outperform the linear methods (Alon et al., 2001)

though there must be a suspicion of overfitting. One study also found that prior seasonal

adjustment of the data can significantly improve forecasting performance of the neural network

model in forecasting market level aggregate retail sales (Kuvulmaz, Usanmaz, and Engin, 2005)

although in wider NN research this conclusion is moot. Despite these claims this evidence of

the forecasting benefits of non-linear models seems weak as we see below.

    Econometric models depend on the successful identification of predictable explanatory

variables compared to the time series model. Bechter and Rutner (1978) compared the

forecasting performance of ARIMA and econometric models designed for US retail sales. They

used two explanatory variables in the economic model: personal income and nonfinancial

                                                                  8
personal wealth as measured by an index of the price of common stocks; past values of retail

sales were also included in alternative models that mixed autoregressive and economic

components. They found that ARIMA forecasts were usually no better and often worse than

forecasts generated by a simple single-equation economic model, and the mixed model had a

better record over the entire 30-month forecast period than any of the other three models. No

ex ante unconditional forecast comparisons have been found. Recently, Aye, Balcilar, Gupta,

and Majumdar (2015) conducted a comprehensive comparative study over 26 (23 single and 3

combination) time series models to forecast South Africa's aggregate retail sales. Unlike the

previous literature on retail sales forecasting, they not only looked at a wide array of linear and

nonlinear models, but also generated multi-step-ahead forecasts using a real-time recursive

estimation scheme over the out-of-sample period. In addition, they considered loss functions

that overweight the forecast error in booms and recessions. They found that no unique model

performed the best across all scenarios. However, combination forecast models, especially the

discounted mean-square forecast error method (Stock and Watson, 2010) which weights current

information more than past, not only produced better forecasts, but were also largely unaffected

by business cycles and time horizons.

     In summary, no research has been found that uses current econometric methods to link

retail sales to macroeconomic variables such as GDP and evaluate their conditional and

unconditional performance compared to time series approaches. The evidence on the

performance of non-linear models is limited with too few series from too few countries and the

comparison with econometric models has not been made.

3.2 Chain level aggregate sales forecasting

     Research at the retail chain level has mainly focused on sales forecasting one year-ahead

(Curtis, Lundholm, and McVay, 2014; Kesavan, Gaur, and Raman, 2010; Osadchiy, Gaur, and

Seshadri, 2013). Accurate forecasts of chain level retail sales (in money terms) are needed for

company financial management and also to aid financial investment decisions in the stocks of

retail chains.

     In general, most of the models used for chain level are similar to those used for market

                                                9
level forecasting (i.e. univariate extrapolation models). However, there are some specially

designed models which have been found to have better performance. Kesavan et al. (2010)

found that inventory and gross margin data can improve forecasting of annual sales at the chain

level in the context of U.S. publicly quoted retailers. They incorporated cost of goods sold,

inventory, and gross margin (the ratio of sales to cost of goods sold) as endogenous variables

in a simultaneous equations model, and showed sales forecasts from this model to be more

accurate than consensus forecasts from equity analysts. Osadchiy et al. (2013) presented a

(highly structured) model to incorporate lagged financial market returns as well as financial

analysts’ forecasts in forecasting firm-level sales for retailers. Their testing indicated that their

method improved upon the accuracy of forecasts generated by equity analysts or time-series

methods. Their use of benchmark methods (in particular a more standard econometric

formulation) was limited. Building on earlier research Curtis et al. (2014) forecast retail chain

sales using publicly available data on the age mix of stores in a retail chain. By distinguishing

between growth in sales-generating units (i.e., new stores) and growth in sales per unit (i.e.,

comparable store growth rates), their forecasts proved significantly more accurate than the

forecasts from models based on estimated rates of mean reversion in total sales as well as

analysts’ forecasts. Internal models of chain sales forecasts should benefit from including

additional confidential variables but no evidence has been found.

3.3 Store level aggregate sales forecasting

     Retailers typically have multiple stores of different formats, serving different customer

segments in different locations. Store sales are dramatically impacted by location, the local

economy and competitive retailers, consumer demographics, own or competitor promotions,

weather, seasons and local events including for example, festivals. Forecasting store sales can

be classified into two categories: (1) forecasting existing store sales for distribution, target

setting and viability, and financial control, and (2) forecasting new store potential sales for site

selection analysis.

     Both univariate time series and regression models are used for forecasting existing store

sales. Steele (1951) reported on the effect of weather on the daily sales of department stores.

                                                 10
Davies (1973) used principal components and factor analysis in a clothing-chain study and

demonstrated how the scores of individual stores on a set of factors may be interpreted to

explain their sales performance levels. Geurts and Kelly (1986) presented a case study of

forecasting department store monthly sales. They considered various factors in their test models

including seasonality, holiday, number of weekend days, local consumer price index, average

weekly earnings, and unemployment rate, etc. They concluded that univariate time series

methods were better than judgment or econometric models at forecasting store sales. At a more

operational level of managing staffing levels, Lam, Vandenbosch, and Pearce (1998) built a

regression model based on daily data which set store sales potential as a function of store traffic

volume, customer type, and customer response to sale force availability: the errors are modelled

as ARIMA processes. However, no convincing evidence was presented on comparative

accuracy. With the rapid changes on the high-street in many countries showing increasing

vacancy rates, these forecasting models will increasingly have a new use: to identify shops to

be closed. We speculate that multivariate time series models including indicator variables (for

the store type), supplemented by local knowledge, should prove useful. But this is research still

to be done.

     Forecasting new store sales potential has been a difficult task, but crucial for the success

of every retailing company. Traditionally, new store sales forecasting approaches could be

classified into three categories: judgmental, analogue regression and space interaction models

(also called gravitational models). Note that any evaluation of new store forecasts needs to take

a potential selection bias into account: candidate new stores with higher forecasts are more

likely to be developed and may see systematically lower sales than forecasted because of

regression to the mean. (The analogue is also true for forecasting new product sales or

promotional sales, see below.)

     The success of the judgmental approach depends on the experience of the location analyst

(Reynolds and Wood, 2010). Retailers often use the so-called “checklist” to systematically

assess the relative value of a site compared to other potential sites in the area. It can deal with

issues that cannot be expressed quantitatively (e.g. access; visibility) and is where intuition and

experience become important. In its simplest form the checklist can act as a good screening tool

                                                11
but is unable to predict turnover. The basic checklist approach can be further developed to

emphasize “some variable points rating” to factors specific to success in particular sectors, for

example, convenience store retailing (Hernandez and Bennison, 2000).

    The analogue regression generates turnover forecasts for a new store by comparing the

proposed site with existing analogous sites, measuring features such as competition (number of

competitors, distance to key competitor, etc.), trading area composition (population size,

average income, the number of households, commute patterns, car ownership, etc.), store

accessibility (cost of parking, distance to parking, distance to bus station, etc.) and store

characteristics (size, format, brand image, product range, opening hours, etc.). Compared with

the judgmental approach, analogue regression models provide a more objective basis for the

manager's decision-making, highlighting the most likely options for new locations. Simkin

(1989) reported the success application of a regression based Store Location Assessment Model

(SLAM) in several of the UK's major retailers. The model was able to account for

approximately 80% of the store turnover, but prediction accuracy for the sales of new stores is

not reported in the paper. Morphet (1991) applied regression to an analysis of the trading

performance of a chain of grocery stores in the England incorporating five competitive and

demographic factors (including population, share of floor space, distance higher order centre,

pull, percentage of married women, etc.). Though the models achieved a high degree of

'explanation' of the variation in store performance, the results on predicted turnover suggested

that the use of regression equations was insufficient to predict the potential performance of

stores in new locations. The pitfalls of regressions may come from statistical overfitting due to

limited data, neglecting consumer perceptions, and inadequate coverage of competition. While

the method can include various demographic variables and is therefore appropriate for retail

operations aiming for a segmented market it is heavily data dependent and therefore of limited

value for a rapidly changing retail environment (as in the UK).

    The spatial interaction model (SIM) (or gravity model) is a widely used sophisticated retail

location analysis tool, which has a long and distinguished history in the fields of geography and

regional science. Based on Reynold and Wood’s (2010) survey of corporate location planning

departments, around two thirds of retail location planning teams (across all sectors) make use

                                               12
of SIM for location planning. Different from analogous regressions which mainly rely on the

data from existing stores in the same chain, SIM uses data from various sources to improve

prediction accuracy: analogous stores, household surveys, geographical information systems,

competition and census data. A spatial interaction model is based on the theory that expenditure

flows and subsequent store revenue are driven by the store’s potential attractiveness and

constrained by distance, with consumers exhibiting a greater likelihood to shop at stores that

are geographically proximate (Newing, Clarke, and Clarke, 2014). The basic example of this

type of model is the Huff trade area model (Huff, 1963). Its popularity and longevity can be

attributed to its conceptual appeal, relative ease of use, and applicability to a wide range of

problems, of which predicting consumer spatial behavior is the most commonly known (Li and

Liu, 2012). The original Huff model has been extended by adding additional components to

make the model more realistic; these include models that can take into account retail chain

image (Stanley and Sewall, 1976), asymmetric competition in retail store formats (Benito,

Gallego, and Kopalle, 2004), store agglomeration effects (Li and Liu, 2012; Picone, Ridley, and

Zandbergen, 2009; Teller and Reutterer, 2008), retail chain internal cannibalization (Beule, Poel,

and Weghe, 2014), and consumer heterogeneity (Newing, et al., 2014). Furthermore, spatial

data mining techniques and GIS simulation have been applied in retail location planning. These

new techniques have proved to outperform the traditional modeling approach with regard to

predictive accuracy (Lv, Bai, Yin, and Dong, 2008; Merino and Ramirez-Nafarrate, 2016).

    Following Newing et al. (2014), let Sij represent the expenditure flowing between zone i

and store j then

                                             W j exp(− β Cij )
                                  Sij = Oi
                                             ∑W
                                              j    j   exp(− β Cij )

Oi is a measure of the demand (or expenditure available in zone i); Cij represents the travel time

between zone i and store j; and Wj accounts for the attractiveness of store j. The attractiveness

term, Wj will itself depend on factors such as accessibility, parking, other store features etc.

Such models are usually validated on in-sample data. But Birkin, Clarke, and Clarke (2010)

criticize this limited approach emphasizing the importance of a hold-out sample (an

unacknowledged reference to the forecasting literature) and show, using DIY chain store data,

                                                  13
that the model can be operationalized with a forecasting accuracy of around 10% (which proved

better than the company’s performance). An important omission is the time horizon over which

the model is assumed to apply, presumably the time horizon of the investment. Birkin et al.

(2010) comment the models are regularly updated at least annually which suggests an implicit

view as to lack of longer-term stability in the models arising from a changing retail environment.

Extensions to the model suffer from problems of data inadequacies but Newing et al. (2014)

argue these can be overcome to include more sophisticated demand terms such as seasonal

fluctuation,and different types of retail consumer with different shopping behaviors.

    Predictive models of store performance are only one element in supporting the location

decision. Wood and Reynolds (2013) discuss how the models are combined with context

specific knowledge and the judgments of location analysts and analogous information to

produce final recommendations. There is no evidence available on the relative importance of

judgmental inputs and model based information. Nor is there much evidence on the accuracy

of the models beyond untested claims as to the model based forecasts being highly accurate

(Wood and Reynolds, 2013) apart from Birkin et al.’s (2010) analysis of a DIY chain. In the

rapidly changing retail environment, we speculate that judgment will again become the

dominant approach to evaluating store potential and store closures. The research question now

becomes what role if any models can usefully play.

    Short-term forecasting of store activity can utilize recently available ‘big’ data in the form

of customer credit (or mobile) transactions to produce shop sales forecasts. The use of the

forecasts a week or so ahead is in staff scheduling. Ma and Fildes (2018) used mobile sales

transactions, aggregated to daily store level for 2000 shops registered on a leading third-party

mobile payment platform in China to show that the forecasts which took into account the overall

activity on the platform (i.e. a multivariate approach) produced using a machine learning

algorithm, outperformed univariate methods including standard benchmarks.

4. Product level demand forecasting in retail

    Product level demand forecasting in retail usually aims to generate forecasts for a large

number of time series over a short forecasting horizon, in contrast to long term forecasting for

                                               14
only one or a few of time series at a more aggregate level. The ability to accurately forecast the

demand for each item sold in each retail store is critical to the survival and growth of a retail

chain because many operational decisions such as pricing, space allocation, availability,

ordering and inventory management for an item are directly related to its demand forecast.

Order decisions need to ensure that the inventory level is not too high, to avoid high inventory

costs, and not too low to avoid stock out and lost sales.

4.1 The hierarchical structure of product level demand forecasting

     In general, given a decision-making question, we then need to characterize the product

 demand forecasting question on three dimensions: the level in the product hierarchy, the

 position in the retail supply chain, and the time granularity (Fig. 5): these are sometimes

 labelled ‘data cubes’..

                    Fig. 5 Multidimensional hierarchies in retail forecasting

Time granularity

     For different managerial decisions, demand forecasts are needed at different time

 granularities. In general, the higher the level of the decision from the operational to the

 strategic, the lengthier the forecasting time granularity. For example, we may need forecasts

 on daily granularity for store replenishment, on a weekly level for DC replenishment,

                                               15
promotion planning, and (initial) allocation planning, while on-line fashion sales may rely on

 an initial estimate of total seasonal sales, updated just once mid-season.

Product aggregation level

    Three levels of the product hierarchy are often used for planning by retailers: SKU level,

 brand level, and category level.

    SKU is the smallest unit for forecasting in retail, which is the basic operational unit for

 planning daily stock replenishment, distribution and, promotion. SKU level forecasts are

 usually conducted across stores up to the chain as a whole and in daily/weekly time steps. The

 number of SKUs in a retail chain may well be huge. E.g., in a supermarket, drugstore or home

 improvement/do it yourself (DIY) retailer today, tens thousands of items need weekly or even

 daily forecasts. Walmart faces the problem of over one billion SKU × Store combinations

 (Seaman, 2018). In a fashion chain such as Zara the number of in-store items by design, colour

 and size can also be of the order of tens of thousands, although forecasting may be conducted

 at the “style” or design level, aggregating historical data across sizes and colours and

 disaggregating using size curves and proportions to arrive at the final SKU forecasts. Online

 assortments are typically far larger, especially in the fashion, DIY or media (books, music,

 movies) business.

    A brand in a product category often includes many variant SKUs with different package

 types, sizes, colors, or flavors. In addition to SKU level promotional planning, brand level

 forecasts are also important where there are cross-brand effects and promotions and ordering

 may be organized by brand.

    However, for many retail decisions, the initial forecasts that are required are more

 aggregate, with a tactical promotional plan being developed across the chain that may well

 take inter-category constraints into account (although whether in practice forecasts have an

 active role in such a plan is an open question). A product category usually contains tens of

 brands or hundreds of SKUs with certain attributes in common, e.g., canned soup, shampoo

 or nails. Categories may be segmented into subcategories, which may be nested in or cut across

 brands. Category level sales forecasting mainly focuses on weekly or monthly forecasts in a

 store, over a chain or over a market, and such forecasts are mainly used for budget planning

                                               16
by so-called category managers, who make large scale budgeting, planning and purchasing

 decisions, which again need to harmonize with the resources needed to actually execute these

 decisions, e.g., shelf space, planograms or specialized infrastructure like available freezer

 space.

     Category management and the assortment decision starts with a category forecast which

 Kök, Fisher, and Vaidyanathan (2015) suggest is based on trend analysis supplemented by

 judgment. The assortment decision on which brands (or SKUs) to exclude as well as which

 new products to add is dependent on the SKU level demand forecasts: the effects on aggregate

 category sales of the product mix depend on the cross-elasticities of the within category SKU

 level demand forecasts, with a long (12 month) time horizon. The associated shelf-allocation

 is, Borin and Farris (1995) claim, insensitive to SKU demand forecast errors.

     In short, whatever the focus, SKU level forecasts as well as their associated own and cross-

 price elasticities are needed to support both operational and tactical decisions.

Supply Chain

    A typical retail supply chain consists of manufacturers, possibly wholesalers or other

intermediaries, retailers’ distribution centers (DCs), and stores in different formats. Retailers

need forecasts for the demands faced by each level in the supply chain. Product-store level

forecasting is often for replenishment, product-DC level forecasting for distribution, product-

chain level forecasting for preordering, brand-chain level for supplier negotiations and

potentially for manufacturing decisions in vertically integrated retailers, such as increasingly

many fashion chains. A key question in retail supply chain forecasting is how to collaborate and

integrate the data from different supply chain levels so that forecasts at different levels of the

supply chain are consistent and provide the required information to each single decision-making

process. From the retailer’s perspective the coordination whilst costly has the potential to

improve availability and lower inventory. It may improve retail forecasting accuracy or service

levels (Wang and Xu, 2014) though some retailers doubt this, apparently only selling rather

than sharing their data. Empirical models analyzing the relationship between POS data and

manufacturing forecast accuracy show improvements are possible though not inevitable

(Hartzel and Wood, 2017;Trapero, Kourentzes, and Fildes, 2012; Williams, Waller, Ahire, and

                                               17
Ferrier, 2014). Empirical evidence on successful retail implementation is limited though

Smaros (2007) using case studies identified some of the barriers and how they might be

overcome (Kaipia, Holmström, Småros, and Rajala, 2017).

4.2 Forecasting within a product hierarchy

    Given a specific retail decision-making question, we first need to determine the

aggregation level for the output of the sales forecasting process. A common option is to choose

a consistent level of aggregation of data and analysis. For example, if one needs to produce

demand forecasts at the SKU-weekly-DC level it might seem ‘‘natural’’ to aggregate sales data

to the SKU-weekly-DC level and analyze them at the same level as well. However, the forecasts

can also be made by two additional forecasting processes within the data hierarchy: (1) the

bottom-up forecasting process and (2) the top-down forecasting process.

    The choice of the appropriate level of aggregation depends on the underlying demand

generation process. Existing researches have shown that the bottom-up approach is needed

when there are large differences in structure across demand time series and underlying drivers

(Orcutt and Edwards, 2010; Zellner and Tobias, 2000; Zotteri and Kalchschmidt, 2007; Zotteri,

Kalchschmidt, and Caniato, 2005). This is particularly true when the demand time series are

driven by item specific time-varied promotions. Foekens, Leeflang, and Wittink (1994) found

that disaggregate models produce higher relative frequencies of statistically significant

promotion effects with magnitudes in the expected ranges. However, in the case of many

homogeneous demand series and small samples, the top-down approach can generate more

accurate forecasts (Jin, Williams, Tokar, and Waller, 2015; Zotteri and Kalchschmidt, 2007;

Zotteri et al., 2005). For instance, different brands of ice cream will have a similar seasonality

with a summer peak, which may not be easily detected for low-volume flavors but can be

estimated at a group level and applied on the product level (Syntetos, Babai, Boylan, Kolassa,

and Nikolopoulos, 2016). Song (2015) suggested that it is beneficial to model and forecast at

the level of data where stronger and more seasonal information can be collected.

    In order to solve the trade-off, cluster analysis has been found useful in improving the

forecast performance (Boylan, Chen, Mohammadipour, and Syntetos, 2014; Chen and Boylan,

                                               18
2007). For example, when aggregating product category level demand over stores, one can

cluster stores according to whether they have similar demand patterns rather than according to

their geographical proximity. A priori clustering based on store characteristics such as size,

range and location is common. Appropriately implemented clustering can enable the capture of

differences among stores (e.g., in terms of price sensitivity) as the clustering procedure groups

stores with similar demand patterns (e.g., with similar reaction to price changes). In these terms,

clustering is capable of resolving the trade-off between aggregate parameterization and

heterogeneity, leading towards more efficient solutions. But so far, the weight of contributions

on this issue focused only on the use of aggregation to estimate seasonality factors (Chen and

Boylan, 2007). These works provided evidence that aggregating correlated time series can be

helpful to better estimate seasonality since it can reduce variability.

     Hyndman, Ahmed, Athanasopoulos, and Shang (2011) proposed a method for optimally

reconciling forecasts of all series in a hierarchy to ensure they add up consistently over the

hierarchy levels. Forecasts on all-time series in the hierarchy are generated separately first and

these separate forecasts are then combined using a linear transformation. So far the approach

has not been examined for retail demand forecasting applications.

     In general, hierarchical forecasting has received significant attention, but most researchers

consider only the aggregation problem for general time series, and have not considered the

characteristics of retail sales data which are affected dramatically by many common factors,

such as events, promotions and weather conditions. Research by Jin et al. (2015) suggests that

for store×SKU demand, in promotional intensive categories, regression based methods

including many of the factors discussed above produce substantially more accurate forecasts.

At higher levels of aggregation, in time and space, time series methods may well be adequate

(Weller, Crone, and Fildes, 2016) though research for retail data remains to be done. But there

is as yet no straightforward answer as to how to generate consistent demand forecasts on

multiple hierarchies over different dimensions.

                                                19
4.3 Product level retail sales data characteristics and the influential drivers of
demand

     At the product level, many factors may affect the characteristics of the observed sales data

and underlying demand. Some of the factors are within the control of retailers (such as pricing

and promotions, and “secondary” effects like interaction or cannibalization effects from listed,

delisted or promoted substitute or complementary products), other factors are not controllable,

but their timing is known (such as sporting events, seasons and holidays), and some factors are

themselves based on forecasts (such as the competition, local and national economy and

weather). There are also many other unexpected drivers of retail sales, such as abnormal events

(like terror attacks or health scares), which manifest themselves as random disturbances to sales

time series which are correlated across category and stores that share common sensitive

characteristics.

     As the result of these diverse effects, product level sales data are characterized by high

volatility and skewness, multiple seasonal cycles, their often large volume, intermittence with

zero sales frequently observed at store level, together with high dimensionality in any

explanatory variable space. In addition, the data are also contaminated by stock-outs where the

consumer is unable to purchase the product desired and instead may shift to another brand or

size or, in the extreme, leave to seek out a related competitor.

Stock-outs: demand vs. sales

     Retail product level demand forecasting usually depends on the SKU sales data typically

captured by POS transactions. However, POS sales data presents an imperfect observation of

true demand due to the demand censoring effect, when the actual demand exceeds the available

inventory. Demand estimates using only sales data would result in a negative bias in demand

estimates of the focal product. At the same time, customers may turn to purchase substitutes

when facing a stock-out in the primary target product: this may increase the sales of substitute

products and result in an overestimate of the substitutes. Academic researchers have long

recognized the need to account for this censoring effect in inventory management. This

literature has been primarily centered on methodologies for dealing with the imperfect demand

                                                20
observations. The methods can be classified into two categories: nonparametric (e.g., Kaplan

and Meier,1958) and parametric models using hazard rate techniques (e.g., Wecker, 1978;

Nahmias, 1994; Agrawal and Smith, 1996). For more detail, see Tan and Karabati (2004) who

provided a review on the estimation of demand distributions with unobservable lost sales for

inventory control. Most of methods are based on stock out events data, while Jain, Rudi, and

Wang (2014) found that stock-out timing could further improve the estimation accuracy

compared with methods based on stock-out events. In the marketing and assortment

management literatures, researchers have focused on the consumers’ substitution seeking

behavior when their target product is facing stock out, which is another way of viewing the

problem of product availability (e.g., Kök and Fisher , 2007; Vulcano, Ryzin, and Ratliff , 2012;

Conlon and Mortimer, 2013).

     Conversely, there is some evidence that at least for some categories, demand depends on

inventory, with higher inventory levels driving higher sales: this has been called a “billboard

effect” (Koschat, 2008; Ton and Raman, 2010). Anecdotally, we have encountered retailers who

know this putative effect as “product pressure”. However, no literature appears to have

leveraged inventories as a driver to improve forecasts.

     The proposed forecasting models in this area are in general explanatory and often require

more information than is readily available, such as periodic stock auditing, customer numbers

and assortment information. In addition, any forecasting algorithm that leverages system

inventory information needs to deal with the fact that system inventories are notoriously

inaccurate (so-called “Inventory Record Inaccuracy” or IRI (Dehoratius and Raman, 2008). As

a consequence models published so far are not suited to forecasting applications. The limited

research reported in the forecasting literature may in part be due the lack of real demand

observations so forecasting accuracy is hard to measure. On the other hand, storing observed

changes in the shelf inventory for every product may be very costly to the retailer, and may not

be adequate to identify every single stock-out instance. Technological solutions may become

more common such as RFID (Bottani, Bertolini, Rizzi, and Romagnoli, 2017). The forecasting

issue is whether out-of-stock positions affect overall service and profitability (within category).

                                                21
Intermittence

        Intermittence is another common characteristic in store POS sales data, especially in

slow moving items at daily SKU level. Fig.6 depicts a SBC (Syntetos, Boylan, and Croston,

2005) categorization (see also Kostenko and Hyndman, 2006) over the daily sales of 1373

household cleaning items from a UK retailer, cross-classified by the coefficient of variation in

demand and the mean period between non-zero sales. 861 items exhibit strong intermittent

characteristics.

   Fig. 6 SBC categorization on 1373 household clean items (Source: UK supermarket data)

     Techniques designed specifically for intermittent demand include Croston’s method

(Croston, 1972), the Syntetos and Boylan method (Syntetos and Boylan, 2001), Levén and

Segerstedt method (Levén and Segerstedt, 2004), Syntetos–Boylan approximation (SBA)

method (Syntetos and Boylan, 2005), and TSB method (Teunter, Syntetos, and Zied Babai,

2011), etc. However, most of these models are tested on demand/ sales time series data from

industries other than retail (e.g., service/spare parts, high-priced capital goods in electronics,

automotive, aerospace and high tech), except for Kolassa (2016), who assessed density

forecasts based on Croston’s method and found them sorely lacking. Also note that while

Croston’s method is intuitively appealing and commonly used in practice – at least as a

                                               22
benchmark –, Shenstone & Hyndman (2005) point out that any possible underlying model will

be inconsistent with the properties of intermittent demands, exhibiting non-integer and/or

negative demands. Nevertheless, Shenstone & Hyndman note that Croston’s point forecasts and

prediction intervals may still be useful.

     As mentioned in the stock-out discussion, POS sales are not the same as the latent demand.

The observed zero sales may either be due to the product’s temporary unavailability (e.g., stock

out or changes in assortment) or intermittent demand. Without product availability information,

it is hard to infer the latent demand using only sales data. Much of the retail forecasting

literatures when dealing with forecasting of slow moving items has not recognized this problem

in their empirical studies (e.g., Cooper, Baron, Levy, Swisher, and Gogos, 1999; Li and Lim ,

2018), while the only exception found is Seeger, Salinas, and Flunkert (2016) who treat demand

in a stock-out period (assuming stock-out is observable) as latent in their Bayesian latent state

model of on-line demand for Amazon products.

     Product level demand in retail is also disturbed by a number of exogenous factors, such as

promotions, special events, seasonalities and weather, etc. (as will be discussed in what follows):

all of these factors make intermittent demand models difficult to be applied to POS sales data.

One possibility is to model these influences on intermittent demands via Poisson or Negative

Binomial regression. Kolassa (2016) found that the best models included only day of week

patterns. One alternative approach, yet to be explored in retail, is the use of time series

aggregation through MAPA (Kourentzes, Petropoulos, and Trapero, 2014) to overcome the

intermittence, which then could be translated into distribution centre loading.

Seasonality

     Retail product sales data have strong seasonality and usually contain multiple seasonal

cycles of different lengths. For example, beer daily sales data shown in 7 exhibit both weekly

and annual cycles. Sales are high during the weekends and low during the weekdays, high in

summer and low in winter, and high around Christmas. Some sales data may also possess

biweekly or monthly (paycheck effects) or even quarterly seasonality, depending on the nature

of the business and business locations. For this reason, models used in forecasting must be able

to handle multiple seasonal patterns. Ramos and Fildes (2018) demonstrate this point, using

                                               23
models with sufficient flexibility but parsimonious complexity to capture the seasonality of

weekly retail data: trigonometric functions prove sufficient.

                      Fig. 7 Beer daily and weekly sales: UK supermarket data

Calendar events

    Retail sales data are strongly affected by some calendar events. These events may include

holidays (Fig.7 shows a significant lift in Christmas, i.e., week 51), festivals, and special

activities (e.g., important sport matches or local activities). For example, Divakar, Ratchford,

and Shankar (2005) found that during holidays the demand for beverages increased

substantially, while other product groups were negatively affected. In addition, SKU × Store

consumption may change due to changes in the localized temporary demographics. Most

research includes dummy variables for the main holidays in their regression models (Cooper et

al., 1999). Certain holidays recur at regular intervals and can thus be modeled as seasonality,

e.g., Christmas or the Fourth of July in the US. Other holidays move around more or less widely

in the (Western style) calendar and are therefore not be captured as seasonality, such as Easter,

Labor Day in the US, or various religious holidays whose date is determined based on non-

Western calendars, such as the Jewish or the Muslim lunar calendars.

Weather

    The demand for some retail products is also strongly affected by temperature and other

weather conditions. For example, there is usually strong support that the sales of soft drinks are

                                                24
higher when the weather is hot (e.g., Cooper et al., 1999; Dubé, 2004). Murray and Muro (2010)

found that as exposure to sunlight increases, consumer spending tends to increase.

Nikolopoulos and Fildes (2013) showed how a brewing company’s simple exponential

smoothing method for in-house retail SKU sales could be adjusted (outside the base statistical

forecasts) to take into account temperature effects.

    Weather effects may well be non-linear. For instance, sales of soft drinks as a function of

temperature will usually be flat for low to medium temperatures, then increase with hotter

weather, but the increase may taper off with extreme heat, when people switch from sugary soft

drinks to straight water. Such effects could in principle be modeled using spline transformations

of temperature.

    One challenge in using weather data to improve retail sales forecasts is that there is a

plethora of weather variables available from weather data providers, from temperatures (mean

temperature during a day, or maximum temperature, or measures in between) to the amount,

duration and type of precipitation, or the sunshine duration, wind speed or wind chill factors,

to even more obscure possibilities. One can either choose some of these variables to include in

the model, or transform them in an appropriate way. For instance, one can define a Boolean

“barbecue predictor”, which is TRUE whenever, say, the temperature exceeds 20 degrees

Celsius and there is less than 20% cloud cover. In addition, there are interactions between the

weather and other predictors, like promotions or the time of year: sunny weather will have a

stronger impact on a promoted ice cream brand than on an unpromoted one, and “barbecue

weather” will have a stronger impact on steak sales at the beginning of the summer, when people

can observe “the first barbecue of the season”, than later in the year after they have been

barbecuing for months.

    Another hurdle is, of course, that weather variables need to be forecasted themselves, in

contrast to intervention variables like prices or promotions that the retailer sets themselves, or

calendar events whose date is known with certainty. This means that weather data can only be

meaningfully used for short-range sales forecasts, since weather forecasts are better than chance

only for a short horizon, or for cleaning past data of historical impacts of, say, heat waves. In

addition, this aspect implies that forecasting exercises that use the actual weather in ex post

                                               25
You can also read